LLM & AI Agents¶

Foundations¶

• transformer architecture - A practical, version-aware guide to attention-based transformer structure, autoregressive decoding, positional information, and production configuration boundaries.
• tokenization - BPE, WordPiece, SentencePiece, context windows, token counting
• embeddings - Vector representations, similarity metrics, embedding models, known issues
• frontier models - GPT, Claude, Llama, Mistral, Gemini comparison and selection guide

Prompting and Generation¶

• prompt engineering - System prompts, few-shot, chain-of-thought, checklist pattern, instruction distillation
• function calling - OpenAI/Anthropic tool use APIs, tool descriptions, validation
• llm api integration - Chat completions, message roles, streaming, parameters, cost management

Retrieval-Augmented Generation¶

• rag pipeline - RAG architecture, hallucination problem, improvement strategies, evaluation
• chunking strategies - Text splitting, chunk sizes, semantic chunking, document loaders
• vector databases - Build vector retrieval around versioned embeddings, authorized metadata filters, provenance, recall evaluation, and safe migration rather than static product rankings.

AI Agents¶

• agent fundamentals - ReAct loop, agent components, types, agent vs workflow
• agent design patterns - Plan-and-execute, reflexion, MRKL, scratchpad, design principles
• multi agent systems - Coordinate multiple agents through task contracts, ownership, authority, state, and evidence boundaries; add agents only when a measured decomposition needs them.
• agent memory - Short/long-term memory, HITL, copilot pattern, conversation management
• agent security - Jailbreaks, prompt injection, data poisoning, defense strategies

Frameworks and Tools¶

• langchain framework - A version-aware guide to LangChain's current agent harness, provider integrations, middleware, state, and production boundaries.
• langgraph - Stateful graphs, conditional routing, human-in-the-loop, multi-agent orchestration
• no code platforms - n8n, FlowWise, Gradio UI building, deployment
• spring ai - Java/Spring Boot LLM integration
• ai coding assistants - Operate AI coding assistants through explicit scope, data, tool, approval, and evidence boundaries instead of product rankings or trust in generated code.
• qwen code - Qwen Code CLI installation, provider contract, and authentication history
• unsloth - Unsloth Core, Studio, Desktop, and model-specific fine-tuning bounds

Model Operations¶

• fine tuning - LoRA, QLoRA, PEFT, OpenAI fine-tuning, data quality
• model optimization - Quantization (GGUF, GPTQ, AWQ), distillation, pruning
• ollama local llms - Local inference setup, quantization levels, model selection
• llmops - Evaluation, monitoring, cost optimization, CI/CD for LLM apps
• production patterns - Deterministic context injection, copilot, workflow decomposition, logging

Additional References¶

• adaptive learning systems - Architecture patterns for AI-powered education systems that adapt to individual learners
• adaptive patterns for autonomous agents - Use explicit task state, bounded hooks, capability-scoped subagents, and evidence-based gates instead of opaque keyword triggers or arbitrary ambiguity scores.
• agent architectures - How to structure the control flow and state management of an LLM agent beyond individual patterns
• agent deployment - Taking agents from prototype to production
• agent evaluation - Evaluate agent behavior with versioned task fixtures, deterministic validators, controlled side-effect checks, and reproducible evidence rather than a single benchmark score.
• agent observability dashboards - Real-time observability for multi-agent and sub-agent systems: hook-based telemetry, event
• agent orchestration - Coordinate model calls, tools, handoffs, approvals, retries, and evidence through explicit task state rather than a framework-specific agent loop.
• agent safety alignment - Build agent safety as explicit authority, data, tool, approval, and evidence boundaries rather than as a prompt-only promise.
• agent scope evasion - Coding agents trained to reduce sycophancy exhibit a documented failure mode: when encountering
• agent self improvement - Techniques for agents to improve their own performance through reflection, step-level reward
• agentic rl competitive programming - GrandCode (2026) achieves grandmaster-level performance on competitive programming problems by
• agentic security 2026 - A threat-model and control guide for tool-using agents, MCP integrations, persistent memory, and irreversible effects. Scope checked 2026-09-03.
• agentic systems landscape 2026 - Multi-agent protocols, SDK comparison, orchestration patterns, and real-world coding agent
• ai adaptive learning systems - A version-aware architecture for learner evidence, deterministic scheduling, constrained LLM tutoring, evaluation, and learner-data safeguards.
• ai agent ide features - Design and evaluate AI-assisted coding environments around workspace isolation, explicit permissions, durable task artifacts, verification, and review.
• autonomous agent evolution - Replacing fixed evolutionary search (agents as stateless workers) with long-lived autonomous agents
• chinese ai coding ecosystem - Chinese AI coding tools, patterns, and community practices: Trae, OpenSpec, MetaGPT, GLM-5
• claude adaptive thinking - Configure and evaluate Claude reasoning effort without relying on fixed, model-specific folklore.
• claude code degradation 2026 - A receipt-based method for diagnosing coding-agent quality, configuration, cost, and availability changes without inventing a vendor incident.
• claude code ecosystem - Use Claude Code plugins, skills, hooks, project instructions, and subagents as explicit, versioned governance surfaces; verify their current schema and effective scope before rollout.
• claude code harness patterns - A practical boundary between instructions, tools, deterministic gates, review, and durable evidence for coding-agent work.
• claude desktop session management - Use supported export, account, and extension controls rather than relying on unversioned local cache internals for conversation recovery or cross-device synchronization.
• claude managed agents - Define organization-managed Claude Code subagents with explicit scope, precedence, tool limits, and verification rather than treating managed configuration as a cloud execution runtime.
• context engineering - Treat model context as a bounded working input and preserve task state, evidence, authority, and retrieval provenance in versioned artifacts rather than fixed token allocations.
• gradio llm interfaces - Rapid prototyping of chat UIs with streaming, markdown rendering, and multi-model comparison
• handoff rollup pattern - How to create a bounded, auditable rollup of long-running agent work without pretending that a summary is lossless.
• kv cache compression - Reducing KV cache memory during LLM inference to enable longer contexts and more concurrent
• llm fine tuning practical - End-to-end guide for frontier API and QLoRA fine-tuning with when-to-use decision framework
• llm persona design and engineering - Design an LLM persona as a versioned behavioral policy with explicit authority, privacy, escalation, and evaluation boundaries rather than as an assumed model personality.
• managed agents - A version-aware guide to Anthropic's managed agent harness: agent configuration, environments, sessions, events, permission policies, and data boundaries.
• multi agent messaging - Inter-agent communication patterns for Claude Code sessions: built-in Agent Teams, hook-based
• multi agent systems architectures 2026 - Multi-agent systems (MAS) have diverged into two primary architectural schools: role-based
• multi session coordination - Durable coordination patterns for several coding-agent sessions: isolated worktrees, manifests, append-only evidence, exclusive-resource leases, and verified integration.
• notebooklm integration - Using Google NotebookLM as a free research backend for Claude Code - token-saving workflows
• oh my claudecode omc architecture - How to adopt the fast-moving OMC plugin without mistaking third-party commands, model routing, or generated state for a stable security or release boundary.
• persona adaptive llm - A decision framework for profile fields, retrieval memory, and adapter-based personalization with tenant isolation, evaluation, consent, and deletion boundaries.
• scaling laws and benchmarks - Chinchilla scaling law, standard benchmarks (ARC, DROP, HellaSwag), and model selection guidelines
• social media mcp tools - A provider-neutral, approval-first design for using MCP to draft, validate, and publish social content without treating a social post as a reversible chat action.
• swarm based review and multisampling in agentic workflows - Generate independent candidates, validate evidence, and select agent outputs through explicit acceptance criteria rather than fixed vote counts or model confidence.
• telegram managed bots - A production-safe guide to Telegram's manager-bot model: creation, token rotation, access settings, state isolation, and lifecycle receipts.
• token optimization - Reducing token consumption in agent systems without degrading task performance
• tool use patterns - How to design, expose, and manage tools for LLM agents
• uml driven agent development - Use small, versioned sequence, state, and trust-boundary diagrams to clarify agent workflows, then validate them in the renderer and CI target that will publish them.